Clash just dropped their 5-hour and weekly rate limits for all users. And right before that announcement, OpenAI released their latest and strongest model: GPT 5.6. That alone would be enough news, but the scoring data around it tells an even more interesting story. Based on the artificial intelligence score index, GPT actually released three new models. You have GPT 5.6 Soul, and then the Terra model and the Luna model. Those correspond pretty closely to what we know from the other side: Fable, Opus, Sonnet. The top five models right now are basically a tug-of-war between Anthropic and GPT. So in the last couple of hours, I've been hammering GPT 5.6 Soul pretty hard, running right into my usage limit, and I've been doing it all inside the GPT desktop app instead of the standalone Codex tool.
If you're wondering why I'm in the GPT desktop app at all instead of Codex, there's a clean reason for it: Codex has officially joined the GPT desktop app. So if you're still trying to use the old Codex application, you're no longer going to see the latest models that GPT offers. You have to download the desktop app to use all the same features you used to have in Codex, plus more. I've been using this desktop app to do encoding work and to build the system that manages my second brain. This article is my honest take on whether this new model is good enough to switch over to, and what the desktop app really gets you day-to-day.
Why the Desktop App Now
When you go to download it, just search GPT desktop app download and install it on Mac or Windows. Right in the dropdown you'll see two options: Work and Codex. Work is sort of like the Claw experience, and Codex is the equivalent of Claw Code inside the desktop. Honestly, I've been using both modes and I can't really tell a feature difference. The sidebar and tool set are the same in both. If I switch to the Codex mode, everything looks identical. All the functionality you count on is there.
Features I Actually Use Every Day
Project Organization That Just Works
Under the project section, you can either create a brand new folder from scratch or link an existing folder on your machine. I have two projects set up right now. One is my ERIC OS, which is my second brain, and the other is my current startup application, bookzero.ai. One project manages all my knowledge, and the other one handles actual application builds. When you expand a project, you see every conversation thread laid out. Click any thread to continue right where you left off. And if you want to branch a conversation, you don't have to remember a slash command like you would in a terminal. There's a fork button. Click it, and you get a fresh new thread with the entire conversation history duplicated. You can rename it, delete it, or just keep going. That kind of flexibility inside a visual UI changes how you explore ideas because there's zero friction to start a variant thread.
Visual Sub-Agent Workflows
One of my recent conversations involved a skill migration where I triggered a dynamic workflow. Inside the GPT desktop app, it spun up multiple sub-agents. Because I was out of credits at the time, the job didn't fully complete, but you can see the architecture: one agent focused on implementation, another on evaluation, and a reviewer agent to check the result. That's very similar to the dynamic workflow you'd build in Claw Code, but here the UI adds something extra. You can click into each agent visually and see exactly what they worked on. A file gets edited, an agent finishes its task, and the final result sits right there in an expandable panel. You get the main conversation thread on the left and the agent details on the right. When you're spinning up multiple agents to do implementation work, this interface gives you a much clearer picture of what's happening in parallel.
Other standard actions are all present: a to-do list you can track, direct output previews, an option to open the code in Finder or your preferred IDE, and an approve-for-me toggle. That approve mode is the equivalent of auto-mode in Claw Code, where it acts on your behalf. Want more control? Hit the plus sign and you can invoke slash-goal, plan modes, call in different agents, or trigger plugins, skills, and MTPs.
Speed Mode and a Credit Warning
At the top, you can switch between models: GPT 5.5, Luna, Terra, Soul, whichever you prefer. You can also change the effort level. And then there's speed mode, which is tempting but expensive. I turned on fast mode for a while and absolutely burned through credits. You'll see your 5-hour usage limit and weekly usage breakdown right inside the app. I've been using the pro plan, and on a single day I went through almost 20% of my total weekly allocation just in a few hours. My recommendation: stick with standard mode. Fast mode gives you results quicker, but it drains credits in a way that isn't worth it for most coding sessions.
There's also a nice scroll bar that lets you quickly scan back through past conversations. It sounds small, but when you have dozens of threads, navigating without it is tedious. Overall, the GPT desktop app gives you a better user experience for agent-assisted coding or second-brain management. And if you have the GPT mobile app, the remote feature still works. You can keep your computer on and control the desktop app from your phone, just like before.
Performance: GPT 5.6 vs. the Field
I've been running this model long enough now to speak from direct experience, even if I can't show you every benchmark on screen. The performance of GPT 5.6 Soul is definitely equivalent to Opus 4.8. It's solid, but it is not up there with Fable 5. Fable 5 is the highest tier, and it's also extremely expensive to run. If you look at the cost side, GPT 5.6 and Opus 4.8 are really similar. The context window and token consumption for input and output are in the same ballpark. Opus 4.8 has a slight edge on latency, but in terms of raw intelligence, coding, and agentic performance, GPT 5.6 scores higher across the index. It's a step up from Opus in the areas that matter most when you're letting the model plan, execute, and review work on its own.
The Best Combo: A Planner and an Executor
What I've found works best is not choosing one model to do everything. You use Fable 5 as the planner and orchestrator, and you use GPT 5.6 Soul as the executioner. That's the combo I've been running, and it's the most effective balance I've tried. The reasoning is straightforward: Fable 5 is the smartest model in the world right now for complex reasoning and planning, but it's pricey. GPT 5.6 is close in intelligence for execution tasks, but at a much lower cost. Just look at the numbers. For Fable 5, input and output combined run around $60. For GPT 5.6, the total is $35. That's nearly half the cost. You don't want to burn Fable credits on straightforward implementation that a slightly less expensive model can handle with high accuracy. And you don't want to trust critical planning logic to a model that might not be the absolute smartest. So you split the work: Fable 5 plans, GPT 5.6 executes. That gets you the highest accuracy results while keeping your costs under control.
How I Actually Apply This
In practice, this means I'll fire up the desktop app, select the right model for the task at hand, and often use multiple threads or agents. In my second-brain project, the initial architecture decisions and knowledge graph design happen with the strongest planning model I can access. Then I switch to the executor model for the granular work, wiring up automations, handling the actual encoding, and managing the sub-agents that do the heavy lifting. The same pattern applies when I'm building products for my startup. I don't try to make one model do everything. I treat them like a team, with each model playing its role according to its strength.
Now, if you're curious about exactly how to combine these models, how to drive the highest accuracy from each output, or how to build agent workflows and AI skills that actually ship, that's what we focus on inside my school community. I built it after years of working as a senior AI software engineer at companies like Amazon and Microsoft. You get a full AI builder roadmap that starts with automations, then moves into AI agents, research systems, SaaS building, and eventually how to productize and market your AI skills. You get the templates, the workflows, weekly live calls every Friday morning, and direct community support. When you get stuck, you ask questions and get help right away. The goal isn't to watch more videos. It's to build, post your progress, get feedback, and start to level up. All the video materials I share on my YouTube channel about Hermes agents, Agentic OS, Claw Skills, Spectrum Developments, and more are also inside the community video library. So if you want to jump in, the first course is waiting, and I'll see you there.
Points clés à retenir
- GPT has released three new models: Soul (5.6), Terra, and Luna, corresponding roughly to Fable, Opus, and Sonnet.
- The old Codex application no longer receives model updates; you must use the GPT desktop app to access the latest features.
- The desktop app gives you a visual UI for forking conversations, project organization, and sub-agent views that terminal tools lack.
- Approve mode and dynamic workflows let you spin up sub-agents that do implementation, evaluation, and review, all with a clickable interface per agent.
- Fast mode dramatically increases credit consumption. Stick with standard mode unless you have a specific reason to burn credits.
- GPT 5.6 Soul performance is roughly equivalent to Opus 4.8, with higher agentic and coding intelligence scores but slightly slower latency.
- The strongest and most cost-effective approach is to use Fable 5 as the planner and GPT 5.6 as the executor, cutting costs by about half while keeping accuracy high.
The GPT 5.6 desktop app isn't just a cosmetic upgrade. It merges what used to be separate tools into one environment where agent workflows have more visual clarity and where you can juggle models according to their actual strengths. The model itself is a serious contender, and pairing it with a top-tier planner ends up being the practical sweet spot. If you're building complex projects right now, that two-model strategy is the one I'd start with today.